Skip to content
wordspace
Home Guide Reading Idea
Download

← Back to home

LLM wiki: understand it in ten minutes

Eight sources, read one by one. What they say, and what changes for you.

You have hundreds of notes and no way to find the thread again. Since April 2026 the answer has a name: instead of rummaging through your notes at every question, you let a model write a wiki and keep it current. Here are the sources that told the story, with a comment of our own and, at the end of each entry, what changes if you use wordspace.

searchevery question re-reads everythingSEARCHat every question???compilecompiled once, then consultedCOMPILEonce, when you savewiki???
Search starts over at every question; a wiki does not. In wordspace it updates when you save, not when you ask.

No excerpts and no long quotes: the links go to the original, which is where it should be read. · Reviewed: 9 · 2026

What it is

GitHub Gist4 · 2026Andrej Karpathy

The person who proposed the pattern

Stop redoing the same work at every question. The model compiles a wiki of markdown files once: the pages cite each other and get richer with every new source. You choose what to feed it and what to ask; the repetitive work goes to the machine.

wordspace does it without you assembling anything: the wiki grows from your notes as you write, and the model runs on your Mac.

Read the original →

Analytics Vidhya4 · 2026Riya Bansal

Why a wiki beats searching

Searching starts from zero at every question. A wiki accumulates. The difference shows when you ask something that lives in five different notes: the first way struggles, the second has already done the work.

That's why wordspace updates the wiki when you save, not when you ask.

Read the original →

themes come from shared labelsa note with no shared label is a group of its ownTHEMES COME FROM SHARED LABELS#memory@elena#structure#travel
Two notes land in the same theme only if they share a real label — a #tag, an @person. A note that shares none is a group of its own, and a valid one.

Putting it to work

AI Maker4 · 2026Wyndo

The same pattern on Obsidian

Someone took the idea and ran it over their own Obsidian vault with an agent: it reads, writes the pages, draws the links. The vault organises itself. The pattern isn't tied to one program.

If your archive is already in Obsidian, the syntax is the same: what you write here is markdown, not an invented format.

Read the original →

DEV Community4 · 2026Kunal

What it costs to build it yourself

An honest report, with the real obstacles: compiling on macOS, preparing the data, a machine brought to its knees. The conclusion is that privacy is paid for in hours of work.

Those are the hours an installed app saves you: here you pick the models from a list, and the numbers are measured on your Mac, not promised.

Read the original →

Where it falls short

GitHub Gist8 · 2026Rohit Ghumare

Four months on: what was missing

Someone who kept a wiki like this running for months names its weak spot: it treats everything written as valid forever. The proposal is to weight claims by how many sources back them, and have new information explicitly supersede the old while keeping the history.

The problem applies here too: which is why the wiki isn't a deposit that only grows, but is rebuilt from your notes when the notes change.

Read the original →

PromptQuorum8 · 2026Hans Kuepper

“Local” isn't enough to keep you safe

A checklist that dismantles a comfortable assumption. Keeping the model on your machine avoids sending your questions to someone else's servers, but it says nothing about the program's telemetry, nor about model files downloaded from anywhere, nor about an interface left open on the network. Design as though the model were already compromised.

It applies to us too: “it runs on your Mac” is a promise to be checked, not believed. The questions on this list are the right ones to put to us.

Read the original →

arXiv4 · 2026Stefan Miteski

An archive that resembles you too closely

What goes wrong when a knowledge base has a single reader: the most-cited pages become the dominant version, and evidence that contradicts them struggles to get in. The author proposes triage, decay and audit, so that contradiction has a route all the way through.

wordspace keeps the link between every claim and the note it came from, so the comparison can always be redone. The objection still stands.

Read the original →

Forte Labs11 · 2023Tiago Forte

The criticism: too many tools, too little problem

Forte tries the generation of apps meant to replace Evernote and turns each one down: too much technique, too much terminology, too slow. The observation cuts against us too, which is why it's here.

How we try not to fall into it: no folders to invent, no system to learn. Capture, and decide later.

Read the original →

The themes, and what wordspace does

Four objections that come from the reading above. Three of them are about us too, and that's what the “what it doesn't do” line is for: without it, this section would be advertising.

Knowledge ages; the pattern doesn't

Raised by: Rohit Ghumare

What wordspace does. The wiki is rebuilt from your notes: correct a note and the page citing it changes with it. There's no parallel archive to keep aligned by hand.

What it doesn't do. We don't track what superseded what. If you change your mind, the new page doesn't say there was an old one, or why it fell.

An archive with a single reader stiffens

Raised by: Stefan Miteski, arXiv

What wordspace does. Contradictions are a feature, not an accident: the app looks for the places where you contradicted yourself and puts them in front of you, with the two notes that don't add up. Answering creates a note of your own, not an automatic correction.

What it doesn't do. Nothing against weight: a theme you've written about a lot counts for more than one barely mentioned, and the wiki doesn't correct that.

“Local” doesn't automatically mean safe

Raised by: Hans Kuepper

What wordspace does. One single network call in the whole app, and it only fires when you import a link yourself. No telemetry, no analytics service, no account: there is no code that sends anything anywhere.

What it doesn't do. The models are downloaded from the internet the first time, and that download goes through Apple's and MLX's libraries: there, the network is theirs, not ours.

The tools weigh more than the problem

Raised by: Tiago Forte

What wordspace does. No folders to invent, no method to learn, no required fields. The Inbox exists to postpone the decision until you have time, which is when decisions get made well.

What it doesn't do. On first launch you pick a model profile, and that is a technical decision we ask of you anyway. We've cut it down to three cards with measured numbers, but it's still a question before you start.

wordspace

A writing app for macOS. Built by Dario Mottadelli.

Product

How it works Guide Reading Idea The study Privacy Download

Legal

Privacy policy Support

Language

itenfrdeespt
This page set no cookies and made no third-party requests. © 2026 Dario Mottadelli. wordspace is a trademark of Dario Mottadelli.